MAGNDATA: Magnetic Structure Database

• MAGNDATA is a curated database of experimentally verified magnetic structures with complete lattice and magnetic-symmetry details essential for condensed-matter research.
• It provides a critical experimental benchmark for validating first-principles, symmetry-based, and machine-learning workflows in magnetic crystallography.
• Researchers use MAGNDATA to assess propagation vector classifications and identify systematic biases in computational magnetic-structure predictions.

MAGNDATA is a magnetic-structure database used in condensed-matter and materials research as a curated reference of experimentally determined magnetic order. In the literature, it is described as the Bilbao Crystallographic Server’s database of magnetic structures, assembled from neutron-diffraction-based determinations, with complete lattice and magnetic-structure information and explicit magnetic-symmetry content. Its principal role has been less that of a broad property table than that of an experimental ground truth for magnetic symmetry, propagation vectors, and ordered spin configurations, especially in studies that benchmark first-principles, symmetry-based, and machine-learning workflows against known magnetic structures (Fahmy, 7 Sep 2025).

1. Definition and institutional context

MAGNDATA is treated in recent literature as a manually curated database of experimentally verified magnetic materials with resolved magnetic structures. One paper describes it as the “Bilbao Magnetic Materials Database / MAGNDATA,” emphasizing that it provides complete lattice and magnetic structure information for each entry and has already enabled downstream discoveries such as magnetic topological insulators (Zhang et al., 2024). Another describes it as a manually curated database of roughly 2,000 magnetic materials with experimentally verified lattice and magnetic-structure information (Itani et al., 2024).

The database is closely associated with the Bilbao Crystallographic Server. In machine-learning work on database correction, MAGNDATA is explicitly described as the Bilbao Crystallographic Server’s database of magnetic structures, containing fully resolved magnetic structures from the 1950s onward and covering both commensurate and incommensurate order (Fahmy, 7 Sep 2025). In symmetry-aided first-principles work, it is further identified as the curated database of experimentally known magnetic structures introduced by Gallego et al. for commensurate magnetic order (Zhou et al., 4 Jan 2026).

This usage establishes MAGNDATA as a reference infrastructure for magnetic crystallography rather than a generic materials repository. A plausible implication is that its value derives from the quality and interpretability of its magnetic-structure annotations, not simply from record count.

2. Scope of stored information

Across the cited studies, MAGNDATA is consistently characterized as a source of experimentally established magnetic structures together with symmetry information. In high-throughput antiferromagnet screening, it is singled out as valuable because it provides experimentally determined magnetic structures with magnetic symmetry information, including both commensurate $(\mathbf q=\mathbf 0)$ and incommensurate $(\mathbf q\neq \mathbf 0)$ cases (Nomoto et al., 2024). In database-construction work, it is described as storing complete lattice and magnetic structure information (Zhang et al., 2024).

The literature also gives several paper-specific size estimates:

Source	Description	Reported scale
(Zhang et al., 2024)	Limited but structurally rich resource	1,890 entries
(Itani et al., 2024)	Manually curated experimental magnetic database	roughly 2,000 magnetic materials
(Fahmy, 7 Sep 2025)	Commensaurate magnetic materials at time of data collection	2,167 commensurate magnetic materials

These counts should be read as study-dependent snapshots or filtered subsets rather than a contradiction. Some papers refer to the overall database, whereas others refer only to the commensurate portion or to the subset matched to a particular computational workflow. This suggests that MAGNDATA’s apparent size depends strongly on the inclusion criteria used by each study.

A recurring technical distinction is the propagation vector. Several later works reduce MAGNDATA labels to a binary problem of zero versus nonzero propagation vector, while others benchmark only the $(\mathbf q=\mathbf 0)$ subset. This reflects the fact that propagation-vector structure is central both to magnetic symmetry classification and to the feasibility of standard first-principles workflows (Fahmy, 7 Sep 2025).

3. Role as an experimental benchmark for magnetic-structure prediction

MAGNDATA’s most prominent modern role is as an external benchmark for computational magnetic-structure prediction. In a high-throughput search for antiferromagnets with anomalous transport signatures, the database is used as an independent experimental check rather than as part of training. The authors state that more than 90% of MAGNDATA structures can be represented by linear combinations of up to three cluster multipole basis elements, and they use MAGNDATA plus published literature to verify 26 out of 28 compounds with partially or fully elucidated experimental magnetic structures (Nomoto et al., 2024).

That study also reports that 34 compounds were identified as “FM-like AFMs,” of which 16 had magnetic structures already established in MAGNDATA. The point of the comparison was not merely structural similarity, but recovery of the experimentally reported magnetic symmetry and zero-propagation-vector ordering patterns relevant to anomalous Hall and spintronics applications (Nomoto et al., 2024). In this setting, MAGNDATA functions as a high-specificity validation target for symmetry-resolved antiferromagnetism.

A second benchmark use appears in symmetry-aided high-throughput calculations based on Landau theory and irreducible representations. There, MAGNDATA provides the experimental reference set against which the workflow is tested. The authors start from 346 MAGNDATA-derived candidate magnetic materials, exclude 39 as inconsistent with Landau’s phase transition theory, and benchmark on 260 materials. Their method correctly identifies the magnetic structure for 207 of these 260 materials, corresponding to 80% accuracy (Zhou et al., 4 Jan 2026).

The significance of these studies is methodological. MAGNDATA is not used merely to confirm that a compound is magnetic; it is used to test whether a workflow recovers the correct ordered state among many symmetry-allowed possibilities. This places the database at the interface of magnetic crystallography, first-principles energetics, and automated materials discovery.

4. Machine learning, label correction, and propagation-vector classification

MAGNDATA has also become important as a training and validation source for machine-learning models intended to improve large computational materials databases. In work on “machine learning magnetism from simple global descriptors,” it is explicitly positioned as the experimentally grounded reference standard against which Materials Project magnetic labels can be assessed (Fahmy, 7 Sep 2025).

In that study, MAGNDATA provides the labels for a binary classification task on the propagation vector: zero versus nonzero. The paper states that MAGNDATA contains 2,167 commensurate magnetic materials at the time of data collection, and that after augmenting MAGNDATA entries with Materials Project descriptors, the resulting MAGNDATA-based learning set has 3,980 entries because some materials map to multiple Materials Project profiles (Fahmy, 7 Sep 2025). Random Forest and XGBoost classifiers trained on these labels achieve validation accuracies of 93% and 92%, respectively, with macro $F_1$ scores of 91% and 90%; test performance remains above 92% with macro $F_1$ above 90% (Fahmy, 7 Sep 2025).

The most consequential use of these models is corrective rather than merely predictive. When applied back to Materials Project ferromagnets, the MAGNDATA-trained models identify a large set of entries likely affected by systematic ferromagnetic bias. The intersection of the Random Forest and XGBoost predictions contains 7,843 materials that are treated as the most confident set of likely misclassified Materials Project ferromagnets (Fahmy, 7 Sep 2025). This makes MAGNDATA a practical instrument for database auditing.

A common misconception is that experimental databases become less useful once larger computational databases exist. The opposite pattern appears here: MAGNDATA’s smaller but experimentally resolved labels make it valuable precisely because they can expose biases introduced by heuristic DFT workflows.

5. Position within the magnetic-database ecosystem

MAGNDATA occupies a distinctive niche relative to newer magnetic-materials databases. Literature-mined resources such as GPTArticleExtractor’s 2,035-entry database and the Northeast Materials Database (NEMAD) with 26,706 magnetic materials are described as broader, more feature-rich, and more directly usable for large-scale machine learning. Yet both works treat MAGNDATA as the closest existing resource when the criterion is resolved magnetic structure rather than bulk scalar properties (Zhang et al., 2024, Itani et al., 2024).

The contrast is explicit. GPTArticleExtractor states that MAGNDATA is limited to 1,890 entries and is therefore too small for some deep-learning applications, but also identifies it as the closest existing resource because it provides complete lattice and magnetic structure information (Zhang et al., 2024). NEMAD similarly positions MAGNDATA as a manually curated database with experimentally verified lattice and magnetic-structure information, while arguing that MAGNDATA is relatively small, has fewer features, and is better suited to single-entry lookups than to data-driven modeling (Itani et al., 2024).

This division of labor is important. MAGNDATA supplies high-fidelity magnetic structures and symmetry labels; larger databases supply broader property coverage, transition temperatures, or machine-learning-oriented feature sets. A plausible implication is that hybrid workflows will remain standard: MAGNDATA for ground truth on ordered states, and broader databases for scale.

6. Scope boundaries, filtering conventions, and naming ambiguity

Studies that use MAGNDATA rarely employ the full database without restriction. Instead, they often impose tight filters to match the assumptions of a computational method. The symmetry-aided Landau-theory workflow, for example, restricts attention to commensurate or $(\mathbf q=0)$ structures, fewer than 40 atoms per unit cell, no fractional occupancy, and materials consistent with its symmetry framework (Zhou et al., 4 Jan 2026). Likewise, high-throughput screening of anomalous-transport antiferromagnets focuses on zero-propagation-vector states and uses MAGNDATA primarily as a validation set for experimentally known structures (Nomoto et al., 2024).

These practices clarify another common misconception: MAGNDATA is not synonymous with “all magnetic materials.” It is a database of experimentally established magnetic structures, and many computational studies further reduce it to subsets compatible with their assumptions. As a result, benchmark accuracy numbers reported against MAGNDATA are always conditional on the subset definition.

There is also a nomenclature issue. An unrelated 2025 proposal for generalized parton distribution analyses introduced an open, YAML-based, GitHub-hosted database also called MAGNDATA, designed to store experimental and lattice-QCD data for GPD phenomenology (Burkert et al., 23 Mar 2025). That usage concerns GPD analyses rather than magnetic structures. In materials and condensed-matter contexts, however, “MAGNDATA” ordinarily denotes the Bilbao magnetic-structure database discussed above.

Taken together, the literature portrays MAGNDATA as a high-value experimental reference for magnetic symmetry and ordered spin structure. Its importance lies less in scale than in its status as an experimentally anchored standard against which high-throughput calculations, symmetry-based workflows, and machine-learning models can be tested, corrected, and interpreted.